Using Random Forest To Select Quasar Candidates In VOICE Data

As luminous Active Galactic Nuclei(AGNs),quasars are crucial to the study of the formation and evolution of galaxies,intergalactic medium,and the large scale structure of our universe.Compared to the traditional spectroscopic observations,multi-wavelength photometric observations are more efficient to select quasar candidates.With the increase of observational depth,however,how to detect faint and high-redshift quasars still remains to be investigated.Many methods and tools have been explored and one of which is machine learning(ML)that has been extensively applied in big data analysis.In this thesis,we use the random forest(RF),one of ML algorithms,to select quasar candidates in VOICE data.For VOICE survey,the limiting magnitudes of single-epoch and final stacked images reach to 24.3 mag and 26.0 mag,respectively.We propose a new method to measure the light curves of the detected objects.With the method,the difference of PSFs between different epochs can be reduced,and the background fluctuations are also suppressed.To apply RF,we derive 14 parameters from the measured light curves and photometric data.Finally,we identify about 1600 new quasar candidates,of which 23 have the probability higher than 0.95,from the parent sample with in total 145296 objects.The median r-band magnitude of these candidates is about 24.0 mag,3.0 mag deeper than that of the SDSS quasar spectroscopic survey.Furthermore,our analyses show that colors and parameters relevant to the light curves(e.g.structure function)are very important to efficiently select quasar candidates.The deficiency of faint and high-redshift quasars in the training sample is a problem for ML based selections,and poses a challenge for near future surveys aiming at selecting quasar candidates through ML techniques.The analyses using the deep VOICE survey data shown in this thesis study provide a useful assessment of the problem and can be beneficial for future research in the field.